Question 480
An assistant must refuse disallowed harmful requests, but candidate models show different safety performance. How should the model be chosen?
Explanation
Correct Answer
Safety suitability needs targeted evidence, not only benign-task fluency or popularity scores.
Incorrect Answers
Validate approved model-provider licensing and permitted commercial usage. That operation would be relevant if: A model is suitable technically but the organization cannot accept its provider's licensing terms. Its corresponding reason or verification is: Legal and procurement suitability are necessary before a model is integrated into a commercial production service. However the problem in this question is: An assistant must refuse disallowed harmful requests, but candidate models show different safety performance. The requirement instead calls for Compare safety evaluations and required policy behavior on representative risky prompts; the question-specific check is: Safety suitability needs targeted evidence, not only benign-task fluency or popularity scores.
Restrict candidate models to compliant deployment locations and permitted data-handling terms. A separate source case supports that approach when: A medical application cannot send patient data to a deployment outside an approved geography, but the fastest model is hosted elsewhere. The supporting rationale or observable result is: Regulatory geography and privacy constraints override convenience even when another model has better raw speed. This question describes a different issue: An assistant must refuse disallowed harmful requests, but candidate models show different safety performance. Address it through Compare safety evaluations and required policy behavior on representative risky prompts, with the required justification or evidence: Safety suitability needs targeted evidence, not only benign-task fluency or popularity scores.
Measure time-to-first-token and streamed output latency for candidate deployments. This is useful for the distinct issue: A voice interface requires very short time to first token, while existing model candidates differ mainly in streaming responsiveness. The relevant technical rationale or evidence for that issue is: Interactive user experience depends on early response timing, not only total completed-response latency. Here the reported issue is: An assistant must refuse disallowed harmful requests, but candidate models show different safety performance. The relevant response is: Compare safety evaluations and required policy behavior on representative risky prompts. The first result cannot verify this different correction: Safety suitability needs targeted evidence, not only benign-task fluency or popularity scores.